The Same TypeScript Costs 73% More Tokens on Claude Than GPT

TLDR Dev · 3 days ago

Anthropic's new tokenizer for Claude models (Sonnet 5, Opus 4.8, Fable 5) produces approximately 30% more tokens from the same code than the previous tokenizer while maintaining the same list price. On identical TypeScript files, Claude's new tokenizer generates 1.73 times more tokens than GPT's o200k tokenizer, making Opus 4.8's effective price approximately $7.50/$37.50 rather than the listed $5.00/$25.00. Users comparing model prices must account for tokenizer differences rather than relying on published per-token rates, as the same work costs 32% more on Claude's newer models despite identical list pricing.